[RFC][PATCH 0/2] Make sparcv8 work again on cas enabled hardware

[Andreas Larsson]

On 2016-11-02 11:05, Torvald Riegel wrote:
> On Tue, 2016-11-01 at 12:51 -0400, David Miller wrote:
>> From: Torvald Riegel <triegel@redhat.com>
>> Date: Tue, 01 Nov 2016 17:46:41 +0100
>>
>>> On Tue, 2016-11-01 at 12:09 -0400, David Miller wrote:
>>>> From: Torvald Riegel <triegel@redhat.com>
>>>> Date: Tue, 01 Nov 2016 16:59:44 +0100
>>>>
>>>>> On Tue, 2016-11-01 at 16:07 +0100, Andreas Larsson wrote:
>>>>>> Any comments are most welcome. The spin lock based sparcv8 semaphore
>>>>>> implementation is currently unchanged by this patchset, but I would say
>>>>>> that that should go as well.
>>>>>
>>>>> Agreed.
>>>>
>>>> I think tossing out all of the ldstub based v8 code is not wise.
>>>>
>>>> I was envisioning adding code to use ldstub on v8 when CAS is not
>>>> available in order to maintain the status quo of what worked and
>>>> was functional before the changes which introduced this problem
>>>> for v8 in the first place.
>>>>
>>>> Having that in place until the kernel-side atomics could be
>>>> implemented, propagated, and supported in glibc would be a nice
>>>> intermediate state compared to what we have now.
>>>
>>> How do you intend to make the synchronization primitives work whose
>>> implementation requires a CAS and for which nobody has provided an
>>> alternative implementation that does not require CAS?
>>
>> The pure userland version will do what has been done for decades,
>> by using a spinlock that protects the word we want to do atomic
>> operations upon.  A hash table of spinlocks is another option.
>
> I know about the available techniques; my question was rather aimed at
> who's going to do the work, in which rough stages, and when.
>
> An external table of locks does not work for process-shared
> synchronization.  Do you plan to not support that, and abort() when
> someone tries to create a process-shared condvar, for example?
>
> Or do you intend to write sparc-specific versions of all the concurrent
> data structures that are process-shared?  Note that in the new condvar,
> for example, there's no unused space in pthread_cond_t that could be
> used for a spinlock.  So you'd have to reorganize quite a bit.
>
> If you want sparc-specific versions, who's going to implement them, and
> when?  What do we do in the meantime?

[resent due to failure on my part to remove standard signature]

Linux kernel commit 8b30ca73b7cc7f2177cfc4e8274d2ebdba328cd5 added
sys_sendmsg and sys_recvmsg to the sys_call_table in
arch/sparc/kernel/systbls_32.S. So sparc32 kernels prior to Linux 4.4 do
not support them as straight up system calls.

Best regards,
Andreas Larsson